ANOVA Analysis of Variance

What is ANOVA and why do we use it?

Can test hypotheses about mean differences between more than 2 samples.
Can also make inferences about the effects of several different IVs, each with several
different levels.
Uses sample data to draw inferences about populations
Looks at the ratio of treatment effect (differences between groups) to error (difference
between individual scores and their group means)

A factor is an independent variable (IV).

o It can be between-group
o Within-subject (or repeated measures)

Mixed designs a bit of both
o
Main effect
o Effect of a factor averaged across all other factors
Interactions
o Effect of a particular combination of factors i.e. 1 factor at a specific level of
another factor.

ANOVA as Regression

It is important to understand that regression and ANOVA are identical approaches

except for the nature of the explanatory variables (IVs).
For example, it is a small step from having three levels of a shade factor (say light,
medium and heavy shade cloths) then carrying out a one-way analysis of variance, to
measuring the light intensity in the three treatments and carrying out a regression with
light intensity as the explanatory variable.
Some experiments combine regression and analysis of variance by fitting a series of
regression lines, one in each of several levels of a given factor (this is called analysis
of covariance ANCOVA).

Well come back to this later.

Why cant you just use several t-tests?

If each time you do a t-test it carries a 5% chance of making a Type 1 Error (i.e.
rejecting H0 when it is actually true: finding an effect when none exists)
Each time you carry out a t-test on the same set of data, the Type 1 Error rate
increases:

Number of Tests
1
2
3
4
5
6

Type 1 Error rate

1 (.95) = 5%
1 (.95)2 = 9.75%
1 (.95)3 = 14.3%
1 (.95)4 = 18.5%
1 (.95)5 = 22.6%
1 (.95)6 = 26.5%

This is the familywise error increased probability of falsely rejecting H0

Using ANOVA allows us make multiple comparisons with the same data, without
increasing the familywise error

Logic: The F-Ratio

ANOVA looks at the variance within data set to find out where it comes from
 Why do scores within a data set vary?
Compares the variance due to the IV with the variance due to individual
differences.
The difference will be significant when the between group variability is
significantly more than the within-group variability.


  

  
  + 
=


If treatment effect = 0, then
=  = 1

=

If treatment effect  , then F 

(The following 3 diagrams have been adapted from Lecture notes on ANOVA1 by Dr Wendy
Johnson)

If all groups are the same then any variance between groups is due to chance the
mean of each group is the same as the grand mean, and you can predict any one
score from the grand mean (of all your samples together)
o H0 cant be rejected; no evidence for differences between the means.

If groups are different then the variance between groups is not due to chance
alone, but a treatment effect (between groups variation) plus error (within group
variation).
You can predict one score from the mean of its individual group
o H0 is rejected in favour of H1 there is a difference between the means of
the groups

To decide which is the better predictor, we divide all the variance into within group
variance (a measure of how much each score differs from its group mean) and
between group variance (how much each score differs from the grand mean)

Steps for one-way ANOVA

1. Calculate SST
Total sum of squares

Based on the best guess for each participant being the grand mean
Total sum of squared differences of each score from the grand mean.
 =  !" !$%&' (
)
+ 1,
 = *$%&'

df T = (N 1)
(i.e. one less than the total sample size)
2. Calculate SSM
Model sum of squares (AKA Treatment sum of squares, SSA).

This is how much variance comes from the fact that data points belong to different
groups
Based on the best guess for each participant being the mean of the group to which
they belong
The difference between each participants predicted score (i.e. the mean of the group
to which it belongs) and the grand mean
a. Calculate difference between each group mean and the grand mean
b. Square each of these differences
c. Multiply each result by the number of participants within that group
d. Add the values for each group together
- =  . !. !$%&' (

/- = 0 1
Where k is the number of groups
3. Calculate SSR
Residual sum of squares (AKA error SS, SSE)

How much variation in the scores cannot be explained by the model?

o Could be caused by individual differences basically anything other than

the experimental manipulation
Can be calculated by:
1 =  Or look at the differences between each score and the mean of the group it
belongs to.
o Distance between each data point and its group mean is squared and
summed.
1 = +!". ! . ,)

Or, using variance (s2):

1 =  *.) +. 1,
i.e. The sum of the variance of each group multiplied by one less than the number
of participants in each group.
/1 = / /= +2 1, +0 1,
=20
i.e. Total sample size (N) minus the number of groups (k)

4. Calculate MSM and MSR

SSM and SSR deal with totals (sums), and are therefore influenced by how many in
each sample
o Need to divide by df before calculating the F ratio
3- =
/31 =
5. Calculate F

1
/1

+/- , /1 , =

331

1-way ANOVA table

Source
Factor

SS
SSM

df
k-1

Error

SSR

k(n-1)

Total

SST

kn-1

MS
MSM=
SSM/k-1
MSR=
SSR/k(n-1)

F
MSM
MSR

Critical F
qf(0.95,dfM,dfR)

6. For higher-way designs

Factorial ANOVA looks at effects of 2 or more IVs on a DV, as well as the effect
of the interaction between them.
e.g. 2-way unrelated design
5%6 = 7 + 8 + 9 + 89 + ":6
5%6 =
/  + ;<% + ;<= + ;<%= "&5%>5"& + //  
Calculate MS

ANOVA Assumptions when can you use an ANOVA?

When parametric assumptions are met:
1. DV is interval or ratio
2. Data is usually in row x column format each row represents a case or participant
and each column a group / category
3. Data comes from a random sample
4. Observations are independent of each other
Shouldnt be able to predict an observation from either other observations
in its group, or from corresponding cases in other groups.
5. Homogeneity of Variance
Variability is equal across all treatment conditions/groups
o Most common violation: floor/ceiling effects; As mean 
floor/ceiling, variability  0
o E.g. Likert scales, Counts, Time/length/volume measurements (>0
by definition)
o Apply Welch correction for violations default in 1-way ANOVA
in R.
Test using Levenes test check for R

o Levene test measures heteroscedasticity violations of

homogeneity of variance
o If significant, then there is some reason to be worried
o p-value will only be approximately correct
o Brown-Forsythe/Welch tests robust tests
ANOVA is robust against this if cell sizes are equal
6. DV is normally distributed
Look at frequency distribution charts
o Check for skew (long tails) floor/ceiling effects, Multimodal
distributions
o Fields: P-P plots of data versus expected
o 1-sample Kolmogorov-Smirnoff test
o Compares empirical cumulative distributions to hypothetical
distribution with same mean and standard deviation
BUT ANOVA is robust if this is violated, if sample size is large, and there
is Homogeneity of Variance
7. Sphericity all the covariances across pairs of levels of the IV are similar
Mauchly's test for sphericity: if significant, the usual F-tables dont apply
Use corrections in the corrected F tables
o Greenhouse-Geisser or Huynh-Feldt probabilities.

Types of ANOVA
1. One-way ANOVA
One independent Variable (IV), explanatory variable or factor, with 3 or more levels
(otherwise youd use a t-test which would give the same result as a 2-level 1-way
ANOVA)
e.g. Comparing the effects of 3 different teaching methods (A, B & C 3 levels of the IV
teaching method) on exam results
Independent Variable Teaching Method
Method A

Method B

Method C

2. Factorial
score= grand mean + main effects + interactions + noise

A factor is an independent variable

2 or more factors each with 2 or more levels each
All combinations of each IV occur
o Between-group / Within-subject (or repeated measures)

o Or both Mixed designs

Random versus fixed factors

e.g. Comparing the effects of different teaching methods (between group variable)
and gender (between-group variable) on exam results
Gender

Method A

Teaching Method
Method B

Method C

Male Method A

Male Method B

Male Method C

Female Method A

Female Method B

Female Method C

Male

Female

Main effects (Green dashed line in table below):

o The effect of a factor averaged across all others

Effect of teaching method, averaged over gender

Effect of gender, averaged over teaching method
o Note that it may be misleading to discuss main effects if there is a
significant interaction.
Interactions / Simple Main Effects (Blue solid line in table below):
o Effect of a particular combination of factors

effect of gender at different levels of teaching method
e.g. is there a difference between male and female under
method A?

effect of teaching method at different levels of gender
e.g. is there a difference between teaching method A,
teaching method B and teaching level C within the males
only?

Gender

Method A

Teaching Method
Method B

Method C

Male Method A

Male Method B

Male Method C

Female Method A

Female Method B

Female Method C

Male

Female

3. Repeated Measures
Each individual is exposed to multiple treatments / multiple levels of an IV

Longitudinal studies usually have repeated measures design testing the same person
over time.
e.g. Each student is tested before, during and after going through each teaching method to
see if theres a difference in exam score.
4. Mixed Design
A combination of within / between design
e.g. Each student is assigned a teaching method (A, B, C), Gender is recorded (Male,
Female) - both between-groups variables and test score before and after administration
of teaching method (Before, After) within-participants variable.
Relationships between factors can be:

CROSSED:
Each level of each factor is encountered at each level of each other factor (see
example above)
NESTED:
2 factors are nested if each level of one occurs only with one level of the other:

e.g. Include year of study

Gender

Method A

Teaching Method
Method B

Method C

Male Method A

Male Method B

Male Method C

Year1

Year 3

Year 5

Male

Year2

Year 4

Year 6

Female
Female Method A
Year1

Year2

Female Method B

Female Method C

Year 3

Year 5

Year 4

Year 6

Steps in an ANOVA
1. Check assumptions
2. Visualise data - graphs
3. Find out if there is any difference between any of the groups - ANOVA
Omnibus Test; H0: 1 = 2 = 3; H1: there is a difference between at least 2 of
the groups not all means are equal.
4. Find out between which groups these differences lie:
Planned comparisons or Post Hoc tests.
o e.g. H0: 2 = 3; H1: 2 3, or directional: H0: 2 = 3; H1: 2 >3
o use multiple t-tests with Bonferonni comparison adjust for
experimentwise error rate
o or Fishers Least Significant Difference (LSD)

o Dunnett 2-tailed
o Tukey-Kramer (HSD) - preferred
o Scheffes

Interpreting and Reporting an ANOVA output

In this example :

Main effect of diet: F(2,36) = 83.52; p < .01

Main effect of supplement: F(3,36) = 17.82; p < .01
No interaction between diet and supplement; F(6,36) = 0.33; p =.917

NB report the exact p-value for a non-significant outcome, and p < .05 or p <.01 for a
significant one.
Planned Comparisons and Post Hoc Tests

Having found a difference, we need to find where the difference lies